Industry-level national-account analysis suggests AI adoption raises measured productivity and reduces input and labor intensity while tilting employment toward younger, less-educated workers; however, estimates depend importantly on how and when AI pervasiveness is dated, highlighting substantial measurement uncertainty.
Early Estimates of the Impact of AI Within BEA’s Industry Economic Accounts
Using BEA industry accounts and proxy measures of AI pervasiveness, the paper finds AI is generally productivity-enhancing and input- and labor-saving and is associated with shifts toward younger, less-educated workers, though key results are sensitive to timing assumptions.
Currently, there is not a line item in the U.S. national accounts that can be used to identify and measure the economic impact of artificial intelligence (AI). Therefore, we use tools to indirectly estimate the impact of AI via the lens of BEA’s industry accounts. Throughout, we discuss important economic measurement challenges and make recommendations for next steps. Our baseline model finds evidence that AI is productivity enhancing and input saving and that AI is associated with a shift toward younger, relatively less educated workers. However, an alternative specification that makes different choices about the timing of the pervasiveness of AI yields less robust results, though it also suggests that AI is labor saving. Our findings highlight the importance of additional research and progress on economic measurement related to AI.
Summary
Main Finding
Using BEA’s industry accounts and Census survey measures of AI use, Highfill and Samuels (BEA WP2026-3) find preliminary evidence that AI is productivity enhancing and input saving at the industry level and is associated with a compositional shift toward younger, relatively less-educated workers. Results are sensitive to modeling choices about the timing and pervasiveness of AI: an alternative timing specification weakens the productivity result but still suggests AI is labor saving. The authors emphasize that measurement limitations make these early inferences tentative and call for substantial improvements in official statistics to better capture AI’s economic impact.
Key Points
- Data signal and timing
- The authors take mid‑2022 as the point when AI entered the broad zeitgeist (e.g., ChatGPT emergence).
- Census surveys show AI use by private firms rose from ~3.4% (2018) to ~5.2% (2022) then to ~4.4% (2023), with “don’t know” responses rising sharply by 2022.
- AI intensity construction
- AI use comes from Census ABS (2019, 2022, 2023) and the 2023 BTOS. Survey questions and definitions vary across years; the study uses a binary indicator (any positive indication of AI use or testing) and classifies industries into quartiles of AI intensity.
- High (AI‑intensive) industries in both 2018 and 2022 include NAICS 51 (information), 53 (real estate & rental), 54 (professional, scientific & technical services), and 55 (management of companies).
- Main empirical approach and outcomes
- Difference‑in‑differences design comparing AI‑intensive (top quartile in 2022) versus other industries, controlling for industry and year fixed effects.
- Outcomes drawn from the BEA‑BLS Integrated Industry‑Level Production Account (ILPA / KLEMS): contributions to real output growth, capital accumulation by asset type, labor (by age and education), intermediate inputs, average labor productivity, and total factor productivity (TFP). The authors study 38 industry‑level variables.
- Main empirical findings (baseline)
- AI‑intensive industries show relative gains in TFP and average labor productivity after 2021.
- AI is associated with input‑saving behavior (lower use of some inputs) and with shifting employment composition toward younger, less‑educated workers.
- Sensitivity and alternative specification
- Changing the assumed timing/pervasiveness of AI reduces the robustness of the productivity effects but continues to indicate labor‑saving effects.
- Limitations emphasized by the authors
- AI‑intensity measures are available consistently only for 19 private‑sector industries (outcomes are available for 61 industries).
- Survey definitions and response categories vary across years; the binary AI indicator counts testing as use and treats “don’t know” as non‑use.
- Measuring AI activity and its prices/inputs (e.g., data, compute, software-as-a-service, free or bundled tools) is challenging; mismeasurement can show up as TFP residuals.
- The analysis is macro/industry‑level and does not address micro‑mechanisms or state‑level heterogeneity.
Data & Methods
- AI exposure / intensity
- Sources: Census Annual Business Survey (ABS) 2019, 2022, 2023 and Business Trends and Outlook Survey (BTOS) 2023.
- Construction: binary indicator of any reported AI use/test; industries grouped by quartiles; AI‑intensive = top quartile in 2022.
- Note: survey wording and definitions of AI vary across years; 2019 and 2023 ABS provide more detailed definitions/examples than 2022 ABS.
- Outcome data
- BEA‑BLS Integrated Industry‑Level Production Account (ILPA / KLEMS) for industry contributions to real output growth, capital by asset type, labor by age and education (composition‑adjusted), intermediate inputs (energy, materials, services), and productivity measures (ALP and TFP).
- Analysis limited to private sector (government excluded).
- Empirical strategy
- Difference‑in‑differences regressions of the form: y_it = α_i + γ_t + β (D_s × Post_t) + ε_it where D_s indicates AI‑intensive industries and Post_t is the post‑(2021/2022) period. Industry and year fixed effects control for level and time shocks; the coefficient β captures relative post‑period changes for AI‑intensive industries.
- Sample windows: pre period up to 2021 (e.g., 1998–2021) and post period 2022–2023; authors also test alternative timing specifications.
- Scope of variables
- 38 industry‑level variables examined; summary tables show notable increases in ALP and TFP for AI‑intensive industries after 2021.
Implications for AI Economics
- Early macro signal consistent with productivity potential
- The baseline estimates provide early macro‑level evidence that AI adoption correlates with higher observed productivity (ALP and TFP) and reduced use of some inputs in AI‑intensive industries. This aligns with the view that AI can be productivity enhancing, but the evidence is preliminary and sensitive to timing choices.
- Labor market composition and distributional concerns
- Findings of a shift toward younger, less‑educated workers in AI‑intensive industries are non‑standard relative to many narratives that emphasize benefits concentrated among highly skilled workers. This suggests heterogeneous pathways: AI may complement some labor types and substitute for others, and these patterns can vary by industry and task content.
- Measurement matters — the IT precedent
- The authors draw an explicit analogy to the 1987–2000 IT “puzzle,” where progress in measurement (hedonic pricing, recognizing software as capital, chain indexes) and time for adoption were crucial to identify IT’s true macroeconomic contributions. The same could be true for AI: improved statistics, new product definitions, and richer capital and price measures may change the measured impact over time.
- Policy and research priorities
- Robust inference about AI’s macroeconomic role requires better, more consistent measures of AI use and AI capital:
- Standardize AI definitions across surveys and track intensity/functional use (not just binary).
- Expand industry coverage and granularity; collect firm‑level and longitudinal AI usage and expenditures.
- Measure AI‑related capital (compute, data centers, specialized software) and service prices/quality changes; consider capitalization rules for AI development (analogous to software).
- Link microdata (firm and worker-level) to industry accounts to trace mechanisms and distributional effects.
- Until measurement gaps are addressed, macro estimates will remain sensitive to sample, timing, and definitional choices; caution is warranted in interpreting early results or using them as a sole basis for policy.
- Robust inference about AI’s macroeconomic role requires better, more consistent measures of AI use and AI capital:
- Overall takeaway
- Highfill and Samuels provide a careful, measurement‑focused early attempt to detect AI’s imprint in official national accounts. Their work suggests promising signs of productivity gains in AI‑intensive industries but highlights that much of the story — magnitude, channels, distributional effects, and persistence — depends on better data and continued research.
Assessment
Paper Typequasi_experimental
Evidence Strengthmedium — Findings are based on nationally consistent BEA industry accounts and show consistent patterns (productivity- and labor-saving) in the baseline, but results are sensitive to alternative timing/specification choices and rely on indirect proxies for AI rather than direct, exogenous variation, leaving room for confounding and measurement bias.
Methods Rigormedium — The paper uses careful industry-account panel analysis and explores alternative specifications and robustness checks, but it lacks a convincing exogenous source of variation (no clear IV or natural experiment), and the analysis depends heavily on proxy construction and timing assumptions, which the authors show materially affect results.
SampleAggregate U.S. industry-level data drawn from BEA industry accounts across multiple years, combined with constructed proxies for AI pervasiveness and linked measures of productivity, input shares, and employment composition by age/education (paper compares a baseline specification and an alternative timing specification).
Themesproductivity labor_markets
IdentificationIndustry-level panel regressions that link constructed measures of AI pervasiveness (proxied using BEA-linked indicators and ancillary proxies) to industry outcomes (productivity, input shares, employment composition), with controls and time/industry fixed effects; identification rests on cross-industry/time variation in these AI exposure measures and robustness checks that vary the timing of AI pervasiveness (no clear exogenous instrument or randomized variation).
GeneralizabilityResults are at industry-aggregate level and may not reflect within-firm or within-industry heterogeneity, U.S.-only national accounts context may not generalize to other countries with different industry structures or measurement systems, Findings depend on proxy-based measures of AI pervasiveness, which may mis-measure true AI adoption, Sensitivity to timing/specification choices limits external validity across different identifications of when AI became pervasive, BEA accounts omit informal activity and firm-level dynamics, restricting inference about micro mechanisms
Claims (7)
| Claim | Direction | Confidence | Outcome | Details |
|---|---|---|---|---|
| Currently, there is not a line item in the U.S. national accounts that can be used to identify and measure the economic impact of artificial intelligence (AI). Governance And Regulation | null_result | high | presence/absence of a national-accounts line item for AI |
0.8
|
| We use tools to indirectly estimate the impact of AI via the lens of BEA’s industry accounts. Governance And Regulation | null_result | high | economic impact of AI (estimated indirectly) |
0.48
|
| Our baseline model finds evidence that AI is productivity enhancing. Firm Productivity | positive | high | productivity |
0.48
|
| Our baseline model finds evidence that AI is input saving. Labor Share | negative | high | use of inputs (e.g., labor/capital inputs) |
0.48
|
| AI is associated with a shift toward younger, relatively less educated workers. Employment | mixed | high | worker composition by age and education |
0.48
|
| An alternative specification that makes different choices about the timing of the pervasiveness of AI yields less robust results, though it also suggests that AI is labor saving. Employment | negative | high | labor use (labor-saving effect) |
0.48
|
| Our findings highlight the importance of additional research and progress on economic measurement related to AI. Governance And Regulation | null_result | high | need for improved economic measurement and further research |
0.08
|